Napari meets Cellpose: Developing Workflows for Finetuning Biological Instance Segmentation Models

Napari meets Cellpose: Developing Workflows for Finetuning Biological Instance Segmentation Models | Ian Coccimiglio | October 28th, Virtual I2K 2024 Authors: Ian Coccimiglio, School of Biomedical Engineering at UBC Brian Northan, True North Intelligent Algorithms Description: In this workshop, I employ Cellpose with napari as a test-case in instructing how to appropriately finetune and evaluate instance segmentation models, with the focus on skeletal muscle images. I use this to create a framework for a robust way of finetuning machine learning models in general. I showcase common pitfalls of machine learning, such as inducing 'catastrophic forgetting', the effect of image normalization, sensitivity of outputs to model parameters, and show how biological segmentation can create chaotic outcomes. Pre-Workshop Instructions: Ideally, have napari and cellpose installed within a conda environment. GPU is helpful, but not explicitly necessary. Keywords: napari, cellpose, python, skeletal-muscle